toktape — published LLM inference runs and their caveats

toktapethe record, not a recording of it

Every run on this page is the file toktape wrote on somebody's own server: the card, the caveats that say when a figure is not quotable, and a replay, all drawn from those same bytes. One word on the machine already running llama-server records yours.

brew install midagedev/tap/toktape other ways to install GitHub

/

Narrow by model, size, engine, quantisation, GPU, host, VRAM, prompt set

Narrowed toos: linux ×Clear all

122 runs, newest first

Qwen3.6-35B-A3B Q6_K · 1 stream · ik_llama.cpp 125 tok/sprefill 2734 tok/s

qwen3.6-35b-a3b

ik_llama.cpp7cd62a3e UD-Q6_K 35B · 3.5B activeP7248 · G5271 · cache 0%ctx 32kfa onrtx-3090 rtx-a6000linux1 stream ! 2 caveats

midagedev

2026-09-20

The README hero: four code-review questions at once 70.9 tok/sprefill 1097 tok/s

qwen3.6-35b-a3b

ik_llama.cpp7cd62a3e UD-Q6_K 35B · 3.5B activeP6065 · G603 · cache 0% · min 597ctx 8kfa onrtx-3090 rtx-a6000linux4 streams ! 4 caveats

midagedev

2026-09-20

Newest first unless you choose another order; there is still no rank column and no score. Every row carries the caveats the card would print, because a result set without them is a leaderboard with the sorting taken out.
toktape publish <run.toktape> puts one here. toktape on GitHub

Page views are counted with Cloudflare Web Analytics: no cookies, no fingerprinting. The toktape binary itself sends nothing anywhere.